The present disclosure relates to a method for classifying a terrain type in an area. It also relates to a system for classifying a terrain type in an area, to a computer program and a computer program product.
Satellites used for providing pictures of the Earth's surface can often generate multi-spectral images, i.e. the images generated by these satellites can comprise information in different wavelength areas, for example from ultraviolet (UV) to infrared (IR). As an example, the satellites WorldView-2 and WorldView-3 operated by the company DigitalGlobe provide images from eight different spectral bands named coastal blue (400-450 nm), blue (450-510 nm), green (510-580 nm), yellow (585-625 nm), red (630-690 nm), red-edge (705-745 nm), near IR (NIR) 1 (770-895 nm), and NIR2 (860-1040 nm).
The images can be analysed for identifying, for example, water, other terrain types, cities, etc. For identifying water, a water index can be generated for every pixel in the images. In one example the water index is defined by defining a ratio
      W    i    WV    =                    ρ                  coastal          ⁢                                          ⁢          blue                    -              ρ                  NIR          ⁢                                          ⁢          2                                    ρ                  coastal          ⁢                                          ⁢          blue                    +              ρ                  NIR          ⁢                                          ⁢          2                    as water index. Here, ρNIR2 denotes the reflectance in the NIR2-spectral band and ρcoastal blue denotes the reflectance in the coastal blue-spectral band. A pre-determined threshold for the water index can then determine whether that pixel should be classified as water or not. One reason why the water index above works is that water usually reflects blue wavelengths quite well, whereas NIR-wavelengths usually are reflected only to small amounts. Another example is
      W    i    GE    =                    ρ        green            -              ρ        NIR                            ρ        green            +              ρ        NIR            defined as a water index. ρgreen denotes the reflectance in the green-spectral band and ρNIR denotes the reflectance in the NIR-spectral band. WiGE can, for example, be used for the GeoEye-1 satellite which has four spectral bands, namely blue, green, red, and near IR (NIR). Also other quantities than the reflectance can be used.
A problem with the existing techniques of classifying areas as containing a certain terrain type, such as water, is that it is difficult to find a threshold which is valid over bigger areas. Typically, it happens that land areas sometimes are mistakenly classified as water. One reason for that is that shadowed areas often can give values for the water index which are on the same side of the threshold as water.
In the US patent application US 2014/0119639 a method for classifying water bodies is presented. First, a normalised difference water index (NDWI) is generated for an area and a segmentation of the area into water-body features and non-water-body features is performed. Then this segmentation is refined by calculating a so called confidence score. This confidence score is calculated via stereo matching of images and denotes how well pixels from different images could be matched together in a stereo matching process. It is assumed that water areas are more difficult to match, which results in that pixels within the water areas in general have much less confidence in a stereo matching procedure than pixels from land areas. A threshold for this confidence can then be determined and pixels originally classified as corresponding to water/non-water can then, depending on which side of the threshold of the confidence score they are, keep or change their status as pixels within/outside water areas.
Performing stereo matching as above puts restraints on the images used in the stereo matching process as they should be taken at the same time of the year for allowing stereo matching, since snow or different appearances of deciduous trees otherwise might make it impossible to find corresponding pixels. Further, stereo matching requires a lot of computational effort.
Although the above examples refer to water, similar problems arise for other kinds of terrain types as well. For these other terrain types similar indices can be defined.